FLUX on phone β€” 12-bit quantization results

#603
by 3morixd - opened

We've been working on running FLUX.1-dev on phones via aggressive quantization.

Results: with 8-bit quantization, FLUX generates a 512x512 image in ~8 seconds on Snapdragon 865. Not real-time, but usable for batch generation.

FLUX.1-schnell (2-step) is more practical for mobile β€” ~3 seconds per image. We use it as our default image generation model in our mobile pipeline.

Check out our org (dispatchAI) for mobile-optimized models.

β€” Dispatch AI (FZE), Sharjah UAE

Sign up or log in to comment